Understanding Global Feature Contributions With Additive Importance Measures

By Ian Covert, Scott M. Lundberg, Su-In Lee. Neural Information Processing Systems, 2020

445 citations recorded by the source corpus.

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Explaining by Removing: A Unified Framework for Model Explanation (2020, methodology): “We follow the approach from Covert et al. (2020), which identifies convergence via the ratio of confidence interval width to the range in values.”
Explaining individual predictions when features are dependent: More accurate approximations to Shapley values (2019, unknown): “A general Shapley framework for global additive importance measures is suggested by [17] The main disadvantage of the Shapley value is that the computational complexity grows exponentially and becomes intractable for more than, say, ten features This has led to approximations like the Shapley…”
Algorithms to estimate Shapley value feature attributions (2022, methodology): “[26], there are several methods that use different coalitional games, including global feature attributions where the value is the model’s mean test loss [77], and local feature attributions where the value is the model’s per-sample loss [16].”
Shapley values for feature selection: The good, the bad, and the axioms (2021, unknown): “Upon review of the literature, we found a number of articles suggesting to use a variant of Algorithm 1 [44], [45], [48], [49], two completely applied uses of Algorithm 1 [46], [47], and one paper critical of the algorithm [53].”